Uncertainty-Aware Recurrent Encoder-Decoder Networks for Vessel Trajectory Prediction

Abstract

In this paper, we propose a deep learning framework for sequence-to-sequence vessel trajectory prediction based on encoder-decoder recurrent neural networks to learn the predictive distribution of maritime patterns from historical Automatic Identification System data and sequentially generate future trajectory estimates given previous observations. Special focus is given on modeling the predictive uncertainty of future estimates arising from the inherent non-deterministic nature of maritime traffic. An attention-based aggregation layer connects the encoder and decoder networks and captures space-time dependencies in sequential data. Experimental results on trajectories from the Danish Maritime Authority dataset demonstrate the effectiveness of the proposed attention-based deep learning model for vessel prediction and show how uncertainty estimates can prove to be extremely informative of the prediction error.